Summary

  • The chip allocation G42 was authorized to import by November 2025 — up to 35,000 Nvidia Blackwell GB300-equivalent chips — is roughly a third of the ~100,000-per-year figure allocated to it in the May 2025 draft agreement, and the draft's 500,000-per-year UAE-wide figure was never a delivered quantity.
  • The regulatory baseline under which all of this sits shifted mid-stream: the AI Diffusion Rule was rescinded in May 2025 before its compliance date, so the tier caps it defined never became the operative constraint; the operative constraint is the security and reporting condition attached to each authorization.
  • Stargate UAE is announced at 1GW within a planned 5GW campus; its first 200MW was under construction for 2026 delivery as of October 2025, and no public source documents it as live.
  • Khazna, G42's datacenter subsidiary, reported 650MW of operational capacity — the clearest public execution datapoint in the group's ledger — against announced additions of over 1GW by 2030.

Reading the ledger, not the announcements

The first useful distinction is between four stages that announcements tend to blur. A draft figure is a negotiating position. An authorization is a permission to buy. An announced cluster is a construction plan. An energized megawatt is capacity that carries load. G42's public record contains all four, and the gap between the largest and smallest figures in that record — 500,000 chips per year in a draft versus roughly 35,000 Blackwell-equivalents authorized — is itself the most informative datapoint in the file.

The draft figure dates from May 2025, when Reuters reported a preliminary US-UAE agreement that would have allowed the UAE to import up to 500,000 advanced Nvidia AI chips per year, with about 20% — roughly 100,000 per year — allocated to G42 and the rest directed to US firms operating in the UAE such as Microsoft and Oracle; the definition of "advanced AI chip" and the security requirements were left to a later working group https://www.reuters.com/business/finance/us-close-letting-uae-import-millions-nvidias-ai-chips-sources-say-2025-05-14/. A month later, Reuters reported, citing five unnamed sources, that the deal was far from finalised: US officials had not yet determined the security conditions or enforcement mechanisms https://www.reuters.com/business/finance/us-uae-multi-billion-dollar-ai-data-campus-deal-far-finalised-sources-say-2025-06-06/. Neither number in the draft was a delivery schedule.

The regulatory ground moved underneath both figures. The Biden-era AI Diffusion Rule, issued on 2025-01-15 with a compliance date of 2025-05-15, had created country tiers and per-country caps — roughly 100,000 H100-equivalents for a Tier-2 country such as the UAE in 2025, rising to 270,000 in 2026 and 320,000 in 2027 — with US firms required to keep at least half of their compute in the United States https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion https://www.rand.org/pubs/perspectives/PEA3776-1.html. In May 2025 the Commerce Department rescinded that rule before it took effect and instructed the Bureau of Industry and Security not to enforce it, while announcing additional chip export-control measures https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens. The consequence for G42's ledger is that the tier-cap framework was never the operative constraint on its deployment; the operative constraints are the conditions attached to individual authorizations.

The authorized figure arrived on 2025-11-19, when Reuters reported that US Commerce had authorized exports of advanced AI chips equivalent to up to 35,000 Nvidia Blackwell GB300 chips each to G42 and Saudi Arabia's Humain, conditioned on what the report described as rigorous security and reporting requirements; Reuters estimated 35,000 Blackwells at roughly $1 billion https://www.reuters.com/business/finance/us-authorizes-export-advanced-american-semiconductors-companies-saudi-uae-2025-11-19/. That is a permission to buy, not a deployment record — and it is roughly a third of the draft's annual allocation to G42, and less than a tenth of the draft's UAE-wide figure.

G42's own announcement the next day framed the approval as enabling "full-scale deployment" and located its compute governance in a Regulated Technology Environment described as approved under US Commerce and BIS guidelines; the company also claimed it operates three Top500 supercomputers, including the region's #2 and #3 systems, plus an announced Maximus-01 system in New York ranked 20th globally, with infrastructure in Abu Dhabi, France and four US states https://www.g42.ai/resources/news/g42-receives-us-approval-advanced-ai-chip-exports-enabling-full-scale-deployment-trusted-ai-infrastructure. These are company claims about its own capacity and standing; the public record examined for this article does not independently verify them.

The datacenter layer: operating versus announced

The clearest execution datapoint in the group's record is at the subsidiary level. WAM, the UAE state news agency, reported that Khazna Data Centers reached 650MW of operational datacenter capacity https://www.wam.ae/en/article/bm78mlz-khazna-data-centres-hits-650mw-operational. Khazna's corporate site positions it as a leading datacenter operator in the Middle East and North Africa https://khaznadatacenters.com/, and its press materials announce additions of over 1GW of hyperscale capacity across multiple countries by 2030 https://khaznadatacenters.com/press-release/khazna-data-centers-to-add-over-1gw-hyperscale-capacity-across-multiple-countries-by-2030/. Reporting on Khazna's corporate structure describes ownership by MGX and Silver Lake within G42's group structure https://www.middleeastainews.com/p/khazna-data-centers-mgx-silverlake. In ledger terms: 650MW energized and carrying load, more than 1GW announced for 2030, and no public document that converts the announced figure into a construction schedule with dates.

Stargate UAE is the flagship announcement and the flagship gap. OpenAI's announcement described a 1GW cluster in Abu Dhabi with 200MW expected live in 2026, with G42, Oracle, NVIDIA, Cisco and SoftBank as partners, developed in coordination with the US government https://openai.com/index/introducing-stargate-uae/. G42's own October 2025 update said the cluster is being developed by Khazna within the planned 5GW UAE-US AI Campus and that the first 200MW had moved from design to construction, with delivery planned for 2026 https://www.prnewswire.com/news-releases/g42-provides-update-on-construction-of-stargate-uae-ai-infrastructure-cluster-302586401.html. Reuters reported the first phase would likely carry 100,000 Nvidia chips at a 10-square-mile site eventually hosting 5GW — a planning estimate, in the same register as the draft chip figures https://www.reuters.com/business/media-telecom/stargate-uae-ai-datacenter-begin-operation-2026-2025-05-22/. The ledger entry for Stargate UAE is therefore: 1GW announced, 200MW under construction, zero megawatts documented as live in the sources examined here.

Around the flagship, the buildout is narrower and more concrete. Microsoft and G42 announced a 200MW datacenter capacity expansion in the UAE in November 2025 https://www.reuters.com/world/middle-east/microsoft-g42-announce-200-mw-data-centre-capacity-expansion-uae-2025-11-05/. Core42, G42's cloud unit, has announced expansion of its AI cloud offering using Nvidia accelerated computing https://www.g42.ai/resources/news/core42-expands-its-ai-cloud-offering-using-nvidia-accelerated-computing. A TOP500 registry entry shows a G42-related system among the listed top systems https://www.top500.org/system/180314/. G42 also announced the Stargate UAE launch alongside a Global Tech Alliance https://www.g42.ai/resources/news/global-tech-alliance-launches-stargate-uae.

The constraint layer

The conditions that pace this buildout are political as much as technical. Bloomberg reported in April 2024 that G42 had agreed with the US to divest from China before its Microsoft deal — a governance condition that preceded the compute approvals https://www.bloomberg.com/news/articles/2024-04-16/g42-made-secret-pact-with-us-to-divest-from-china-before-microsoft-deal. Axios reported in December 2024 on the US-UAE Microsoft-G42 AI chip arrangements that formed the corridor's early structure https://www.axios.com/2024/12/07/us-uae-microsoft-g42-ai-chips. CSIS analysis places the UAE's AI ambitions in the context of compute access and governance in US-UAE technology relations https://www.csis.org/analysis/united-arab-emirates-ai-ambitions. The Federal Register rule text, though rescinded, remains the best public articulation of the control logic — validated end users, model-weight controls, compute caps — that any replacement framework will echo https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion.